Qwen 3.5 Small — Apache Multimodal Edge From 0.8B to 9B
Alibaba Qwen completed Qwen 3.5 rollout with Small series (March 2): dense 0.8B/2B/4B/9B open weights, native text/image/video in one stack, Apache 2.0 — on-device/edge after mid-February 397B MoE flagship. Competitive-vs-larger benches are vendor/community — not Artificial Analysis locks.
Corrected October 14, 2026: Re-dated from March 1 to March 2, the date Qwen's GitHub release log gives for the availability of the four Qwen 3.5 Small models.
U.S. labs chase cluster Elo; Alibaba’s proprietary close of the Qwen 3.5 family is ship the edge SKUs — native multimodal dense models small enough for phones and laptops under Apache 2.0.
Alibaba Qwen released the Qwen 3.5 Small series (March 2): four dense open-weight models — Qwen3.5-0.8B, 2B, 4B, 9B — aimed at local inference. Natively multimodal through early-fusion vision-language training. Builds on larger Qwen 3.5 foundation that began with a 397B MoE flagship in mid-February. Reports: 9B competitive vs much larger models on some targeted benches (community/vendor). License: Apache 2.0. Context: qwen.ai/blog family announcements.
Claims vs checks#
Model sizes, Apache 2.0, and on-device positioning are Qwen/Alibaba primary plus early-March coverage. Bench “competitive with larger” claims are vendor/community; independent rankings were not available at the time of writing.
Limits#
- On-device quality depends on quantization and NPU/GPU path.
- Small multimodal still hallucinates; offline ≠ private if apps phone home.
- The MoE flagship and the Small series serve different buyers.
Sources#
- Qwen blog / Qwen3.5 announcement context.
- QwenLM/Qwen3.5 on GitHub: release log entry dated 2026-03-02.
- Alibaba Qwen announcements on model family rollout.
- Benchmark and deployment analysis from developer communities (March 2026).
- Hugging Face: Qwen/Qwen3.5-9B model card